text-super-resolution

The problem

This repository addresses the problem of generating high-resolution images based on low-resolution ones. The dataset contains images of text. Full solution with step-by-step approach can be found in Project3.ipynb file.

Bibliography

1. TextZoom Dataset: https://paperswithcode.com/dataset/textzoom, https://github.com/JasonBoy1/TextZoom, https://arxiv.org/pdf/2005.03341v3.pdf
2. NEOCR Dataset: http://www.iapr-tc11.org/dataset/NEOCR/neocr_metadata_doc.pdf
3. Super-resolution: https://arxiv.org/pdf/2103.02368v1.pdf
4. Models:
   • SRResNet: https://arxiv.org/pdf/1609.04802.pdf
   • Text Gestalt: https://arxiv.org/pdf/2112.08171v1.pdf, https://arxiv.org/pdf/1706.03762v5.pdf
   • ESRGAN: https://arxiv.org/pdf/1809.00219.pdf

How to use the notebook

1. Download the data first and paste it into /data directory
2. In order to use neptune.ai you need to provide your api token (paste your token in ./cfg/tokens/api_token.yaml file in the format token: <your-api-token>).
3. Prepare environment using one of two options:
   • Install dependencies using pip install requirements.txt and set PYTHONPATH=src
   • Use Docker:
     • docker build -t <image-name> .
     • docker run -p 8888:8888 <image-name>

Runtime environment

We trained our models with cuda, having two independent gpus available:

• NVIDIA GeForce GTX 1060 6GB
• NVIDIA GeForce RTX 2080 Ti 11GB